An AI agent researching a company can read its website and still miss the information needed to make a useful decision: financial records, recent hiring, regulatory filings or the people responsible for buying software. Firecrawl is betting that connecting agents directly to those sources—and paying information suppliers—will make the software more useful at work.
Today, the firm is launching a new service, Alexandria — named for the famed Ancient Greco-Egyptian "Great Library" that was destroyed by the forces of Julius Caesar — brings web retrieval, specialized indexes and external data providers into one connection that developers can add to their applications or people can use through AI assistants. The company says the catalog now includes more than 550 providers, spanning financial information, scientific research, public records, software documentation and business contacts.
Firecrawl is also releasing a white paper that argues enterprises should give agents complete assignments, access to authoritative sources and explicit checks on their work.
The report builds the business case for Alexandria, which the company announced alongside a $75 million Series B on September 22. Smash Capital led that round, with participation from Altos Ventures, Nexus Venture Partners, Y Combinator, Freestyle and Offline Ventures.
For enterprise buyers, the proposition is practical: reduce the engineering and procurement work involved in connecting many information sources, while giving agents material they cannot reliably reconstruct from their training.
Customers access Alexandria through Firecrawl’s hosted service, API or Model Context Protocol integration. The materials reviewed for this article do not establish an open-source release of Alexandria itself; the catalog combines commercial sources with publicly accessible information governed by its own terms.
“We're moving into a world where the cost of generating lies is going to zero, right?” Firecrawl CEO and co-founder Caleb Peffer told VentureBeat in an interview.
“But what that also means is that if you hold the truth, if you hold valuable verified information, that's valuable.”
Beyond scraping
Firecrawl’s original utility centers on turning website content into material AI systems can readily process. Alexandria expands that role: an agent can discover an appropriate source, inspect its capabilities and retrieve information through Firecrawl, rather than requiring a developer to build and maintain every connection separately.
The distinction matters when a task crosses several domains. Evaluating a prospective customer might require company information, funding history, employee records and website traffic. Finding those sources is only the first step; each can have different authentication, billing and response formats.
Firecrawl’s lead-enrichment example illustrates the intended workflow. An assistant uses services including FullEnrich, Apollo, Fireround and Similarweb to turn inbound email addresses into company and contact profiles, then assess whether each lead fits a business’s criteria. The instructions are free, but the underlying data calls carry their listed credit charges. Some providers require a team administrator to accept their terms before use.
In the interview, Peffer confirmed that users can give an agent access to the broader library and let it select appropriate sources and spend credits, rather than manually choosing a separate integration for every question.
“The way I see this is our end customer is a user's agents,” he said.
That framing puts Firecrawl between the model and the information the model needs during a task. Users still rely on their chosen assistant to interpret the retrieved material and produce the requested output.
How agents find and run tools
Firecrawl’s Alexandria documentation describes a three-step process: discover a relevant tool, inspect its required inputs and expected response, then execute it. Developers can access that workflow through JavaScript and Python SDKs, HTTP requests, a command-line interface or MCP. A combined search returns both website results and ranked tools, identifying each tool’s provider and capability.
Developers use findTools or find_tools to examine a tool’s inputs, output format and price. Execution runs through Firecrawl’s existing scrape interface: the request supplies a provider, capability and options under an alexandria field, without requiring a webpage URL. A podcast search, for example, can call Particle’s episode-search capability with a topic and result limit. Applications must check individual tool results for errors before using the returned data.
Firecrawl’s own Agent endpoint requires a separate opt-in. API and SDK requests must include exchange: { "enabled": true } to let Agent use the team’s Alexandria providers during research; otherwise, those runs use only the web. Developers can also pin up to five providers through toolkits. Connecting a library therefore does not mean every Firecrawl Agent request automatically uses it.
What the new 'library' of Alexandria contains
The Alexandria catalog mixes commercial information products, public databases and Firecrawl’s own indexes. Screenshots of the service reviewed by VentureBeat show both the scope of those sources and the differences in what customers buy.
Benzinga provides financial news and analyst-related information. Fiscal.ai supplies company financials and related records. Particle offers podcast search, speaker-attributed transcripts and other media analysis. Developer-oriented entries include GitHub, Hugging Face, Artificial Analysis and Firecrawl’s Developer Index.
Other entries draw on public scientific and government resources. The catalog includes ClinicalTrials.gov, ClinVar and the AlphaFold database, alongside local government records and public economic information. Retail and consumer-facing entries extend to Shopify product catalogs, app stores and restaurant information.
The catalog is therefore broader than a collection of exclusive, privately licensed datasets. A substantial part of the offering packages information that already exists publicly into functions an agent can call. The potential benefit lies in consistent access, source discovery and reduced integration work; a listing alone does not establish an exclusive partnership with the organization whose information appears there.
Firecrawl’s own collections add another layer. Its Government & Legal Index describes more than 30 million pages from U.S. government sources, including laws, court materials, permits and agency guidance. Its Research Index spans scientific metadata, abstracts, citation relationships and relevant full-text passages. That description should not be read as unrestricted access to the complete text of every indexed paper.
The entries also reveal useful boundaries. The GitHub listing says it excludes email and contact data. The Europe PMC entry distinguishes access to article records from access to full text permitted for commercial reuse. Such distinctions matter for companies that intend to incorporate retrieved information into customer-facing products.
Paying for knowledge
Firecrawl’s strategy combines a distribution service for information providers with a shared purchasing mechanism for agent users.
Peffer confirmed in the interview that customers spend Firecrawl credits and suppliers receive payment. He cited Benzinga as an example; the company’s September announcement also identifies a paid agreement with Wikimedia Enterprise for direct access to Wikipedia information. Laura Franzese, Firecrawl’s head of brand, said larger suppliers have multiyear contracts, but declined to disclose detailed commercial arrangements.
“It depends on the value of the data and at the end of the day that's just a supply and demand problem,” Peffer said when asked how Firecrawl determines prices.
The company has also outlined plans to let more individuals and organizations earn from their knowledge. Its September announcement described a forthcoming self-service contribution system. That ambition extends beyond the established provider agreements, and should not be confused with a demonstrated, generally available marketplace where anyone can immediately upload expertise and collect revenue.
The economic argument is that agents could become recurring customers for specialized information. A useful dataset would earn revenue as software retrieves it to complete assignments. For buyers, however, the same automation makes spending controls important: an assignment can generate several paid lookups before the assistant returns an answer.
What access costs
Firecrawl’s published pricing starts with 1,000 free credits each month. Monthly subscriptions include Hobby at $19 for 5,000 credits, Standard at $99 for 100,000, Growth at $399 for 500,000 and Scale at $749 for one million. Annual billing lowers the advertised monthly equivalents to $16, $83, $333 and $599, respectively. Enterprise pricing is negotiated.
The documentation separates discovery from execution: finding Alexandria tools and inspecting their inputs and prices costs nothing; running them incurs the listed charge. A tools-only search is free, while combining tool discovery with web results retains the normal web-search charges. Matching tools to a website is also free, but the underlying web search or page scrape remains billable.
Alexandria’s provider cards show why a credit balance does not translate into a fixed number of completed assignments. The November 19 screenshots display the following rates:
Source or capability | Credits displayed |
|---|---|
Wikipedia | 0–1 |
GitHub | 1–5 |
Benzinga | 0–45 |
Fiscal.ai | 0–30 |
Particle | 5–50 |
Firecrawl Developer Index | 2 per 10 results |
Firecrawl Government & Legal Index | 0 |
Firecrawl Research Index | 0 |
These are displayed provider ranges or capability rates, not a promise that every operation returns the same amount of information. Several cards also show provider-specific trial allocations. Buyers need the applicable operation’s billing unit and their account’s terms to calculate a reliable task cost.
The free Government and Research Index tools still have usage limits. Calls through Alexandria share the per-minute and daily allowances of the indexes’ own endpoints, so acvvcessing an index through Alexandria does not create a separate quota. Daily allowances depend on the subscription plan; Scale and Enterprise remove those daily caps, while per-minute limits still apply, according to the index rate-limit guidance referenced in the documentation.
Provider permissions also constrain execution. When a supplier requires additional terms, Alexandria returns an error directing an organization administrator to review and accept them. The documentation requires explicit user authorization before an agent accepts terms. Firecrawl Agent runs skip providers whose terms remain unaccepted and identify them in exchange.skippedProviders, giving developers a way to detect missing sources rather than assume the entire catalog was available.
Firecrawl offers paid-plan credit replenishment with a configurable monthly limit. Its enterprise plan lists single sign-on, service commitments and zero retention of scraped content for approved uses. Those features warrant checking against the specific sources and workloads a company plans to use; the pricing page does not establish identical retention terms across every external provider.
What the benchmarks show
Firecrawl’s new white paper released today, AI’s Search-Bar Problem, links better access to information with better agent output.
Its strongest numerical evidence concerns retrieval and answer quality, rather than measured financial returns.
In one internal evaluation involving 581 requests for official U.S. legal sources, an agent relying on memory placed the official document first 47.8% of the time. Firecrawl’s Government & Legal Index achieved 76.2%, an increase of 28.4 percentage points, or roughly 1.6 times the baseline. The white paper reports that the index placed the official document among its first 10 results in 96.6% of cases.
Those results measure whether the system locates the relevant source. They do not establish that it interprets a statute correctly, completes an entire legal assignment or outperforms every competing retrieval service.
A separate evaluation in the Alexandria announcement reports answer-quality scores 21% higher across 845 tasks than agents using built-in web tools. Firecrawl says it used the same model and prompts and evaluated answers through blind AI judging.
The distinction between the tests matters: the legal comparison uses a memory-based baseline, while the broader evaluation compares against web tools. Neither figure represents a demonstrated 21% increase in employee productivity, and both come from the vendor selling the service.
The white paper recommends auditing how teams use AI, defining complete assignments, identifying necessary sources, requiring supporting citations and measuring completion time and rework. That offers companies a way to test the proposition against their own workflows, rather than treating benchmark scores as a substitute for deployment results.
Exa is a closer rival
Alexandria sits within a widening market for services that supply information to AI systems. The competitive comparison depends on when and how that information gets used.
Mercor describes its work as organizing domain experts to evaluate models, identify weaknesses and create data for further training. micro1’s Realm offering similarly centers on expert-generated datasets, evaluations and reinforcement-learning environments. Both compete for valuable human knowledge and enterprise attention, but those offerings address a different purchase from retrieving a source during an agent’s assignment. Neither cited page provides a directly comparable per-lookup price.
Peffer drew that distinction when asked about those businesses: “our main customers are not big labs who are buying data in bulk and training on it.”
A closer comparison is Exa Connect, which also connects agents with external information providers. Its self-service sources include Particle, Similarweb, Financial Datasets and others; additional services such as Crunchbase and ZoomInfo require access by request. Exa advertises provider-call charges of $0.005 to $0.03, added to its Agent pricing.
That overlap challenges any suggestion that Firecrawl alone has discovered the opportunity to combine search with paid data. Exa also lets customers bring existing provider credentials, an option relevant to enterprises that already have data contracts.
Exa’s search pricing separately lists Lite, Auto and Deep search at $4, $7 and $12 per 1,000 requests, respectively, with up to 10 results included. These search prices should not be compared directly with an Alexandria provider card without matching the underlying operation and result volume.
Tavily offers another alternative for retrieving and preparing web information, with search, extraction, mapping, crawling and research APIs. Its pay-as-you-go rate is $0.008 per credit: a basic search consumes one credit and an advanced search two, equivalent to $8 and $16 per 1,000 searches before discounts. Its documented pricing emphasizes those retrieval operations rather than Alexandria’s catalog of separately priced suppliers.
The strategic comparison is consequently about source coverage, integration effort and total task cost as much as search price. Alexandria’s breadth and combination of public and paid sources give buyers reasons to evaluate it. Exa’s overlapping provider network means catalog size alone cannot establish a durable advantage. The evidence reviewed here does not support a universal winner on accuracy, latency or cost.
Reaching agents inside sandboxes
Firecrawl is also working on access from restricted execution environments. Its November 18 Docker Sandboxes integration lets coding agents retrieve web information through Firecrawl while keeping the Firecrawl API key outside the sandbox.
The integration installs a pinned Python SDK and adds permission to reach api.firecrawl.dev. A host-side proxy replaces a placeholder credential with the real key when it sends authorized requests. Firecrawl’s servers retrieve the requested pages, so the agent does not need direct network permission for every destination website.
This can help a coding agent consult current documentation or release notes from within a controlled environment. The integration adds to the existing network policy, however; it does not automatically restrict all other traffic to Firecrawl. Docker kits remain in Early Access, and Firecrawl labels its kit experimental.
The Docker post describes Alexandria access as requiring beta-enabled keys, while the Alexandria documentation supplied for this article presents API, SDK, CLI and MCP access without stating a general beta-key requirement. The materials do not resolve that discrepancy, so buyers should confirm eligibility for the Docker integration rather than assume the entire Alexandria service requires beta approval.
The enterprise test
Firecrawl reports more than 1.5 million users and 150,000 companies using its services, including teams at 67 Fortune 100 companies.
Those company-wide figures provide a potential distribution base for Alexandria, but they do not establish how many organizations have adopted the newer service or how extensively they use it.
The remaining deployment questions are concrete. How fresh is each source? What rights travel with retrieved information? How does a team constrain an agent’s spending? Can reviewers trace an answer back to the precise record the agent used? Peffer did not provide a detailed freshness schedule in the interview, and the catalog’s varied sources make a single blanket guarantee insufficient.
Firecrawl’s opportunity is to make acquiring useful information a routine part of an agent’s work. Its challenge is to show that the combined cost of retrieval, model use and human review buys a better completed assignment. For enterprises, that is the result worth measuring.
